EnterpriseClawBench - Full Set (DeepAgents): leaderboard
EnterpriseClawBench (Zhong et al., Horizon Research and Frontis.AI, 2026) builds agent tasks from real enterprise workplace sessions: the agent reads uploaded files, uses tools and must deliver business artifacts such as documents, spreadsheets, slides or web pages, which hard rules and LLM judges score. This board is the paper's run on the full 852-task set in the DeepAgents harness, a task score in percent (higher is better); the per-harness leaderboards cover the audited 120-task Lite subset.
Metric: Task Score (%; full 852-task set, DeepAgents harness). Source: arxiv.org. Saturation forecast: Around February 2027. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 76.6 | #48 |
| 2 | Claude Sonnet 4.6 | 74.9 | #158 |
| 3 | Claude Haiku 4.5 | 63.2 | #568 |
| 4 | GPT-4.1 Mini | 33.6 | #775 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=enterpriseclawbench-full-set-deepagents · How It Works · Data refreshed daily, snapshot 2026-11-12.